[Misc]feat: adapt to vLLM main (0d4d334e...1c8e9c03)

[Meihan-chen]

main2main Final Summary: vLLM Ascend Upgrade

Overview

Successfully upgraded vllm-ascend to track upstream vLLM commit 1c8e9c0399f6a6a98f406dce5947a2ad318e195a.

Commit Range

• Base commit: 0d4d334eaa583b9c09aa4eb7538c22db99fd84b3
• Target commit: 1c8e9c0399f6a6a98f406dce5947a2ad318e195a
• Total upstream commits: 52
• Steps processed: 10
• vllm-ascend commits created: 10

Step-by-Step Summary

Step	Upstream Range	Commits	Adaptation	CI Status
1	0d4d334..27b85d2	6	No-op	Passed
2	27b85d2..2676ab1	1	No-op	Passed
3	2676ab1..4b364f8	2	No-op	Passed
4	4b364f8..75fd68c	1	No-op	Passed
5	75fd68c..491e8d8	10	No-op	Passed
6	491e8d8..bd9dbe6	13	No-op	Passed
7	bd9dbe6..787bc0d	12	No-op	Passed
8	787bc0d..36e74c9	1	No-op	Passed
9	36e74c9..0867497	1	No-op	Passed
10	0867497..1c8e9c0	5	No-op	Passed

Key Upstream Changes Absorbed

Models

• InternS2 Preview - New multimodal model support
• DeepSeekV4-Pro - CUDA graph full and piecewise mode
• Qwen3.5-MTP - Weight loading fix with runai_streamer
• Gemma3/4 - Multi-GPU bugfixes, CPU compilation

Core

• Thinking budget bugfix - Fixed inverted condition in gpu_input_batch.py
• KV cache coordinator - Optional alignment_tokens parameter (backward compatible)
• SWA block caching - Skip blocks that can never serve prefix-cache hit

Attention

• MLA optimization - Remove additional copy in _v_up_proj bmm
• FlashAttn FP8 - Fix supports_kv_cache_dtype for FP8 variants
• RISC-V RVV - New CPU attention backend

LoRA

• Layer replacement - Use maybe_get_oot_by_class for out-of-tree layers
• LoRA dedup - Deduplicate wrapping for modules with multiple paths

Infrastructure

• FlashInfer bump - v0.6.11.post2
• lm-eval bump - 0.4.12
• CUDA 12.9 builds - PyTorch manylinux_2_28 base
• Mooncake disk offloading - KV connector feature

Security

• v2 endpoints - API key authorization

Why No Adaptation Was Needed

1. Hardware-specific changes: Most NVIDIA (CUDA, FlashInfer) and ROCm changes don't affect Ascend NPU
2. Upstream registry: Model additions use upstream's model registry which vllm-ascend inherits
3. Own implementations: vllm-ascend has its own attention kernels, MoE implementations, KV connectors
4. Backward compatible APIs: New optional parameters with defaults don't break existing calls
5. Internal changes: Many changes are in upstream's internal implementations that vllm-ascend doesn't override

Files Updated in vllm-ascend

All 10 commits updated the same files (commit reference updates):

• .github/workflows/dockerfiles/Dockerfile.lint
• .github/workflows/pr_test_full.yaml
• .github/workflows/pr_test_light.yaml
• docs/source/conf.py - main_vllm_commit updated each step

CI Verification

All 10 steps passed e2e-singlecard-light CI:

• No code bugs
• No test failures
• All commit references validated

Final State

• vllm-ascend HEAD: 80b1005233cff30f28e970e932b34bc9381d9171
• vllm upstream checkout: 1c8e9c0399f6a6a98f406dce5947a2ad318e195a (target achieved)
• Ready for: User review, optional push to remote

Recommendations for Reviewer

1. Verify commit messages follow project conventions (Signed-off-by)
2. Confirm no temporary files left in repo
3. Review step summaries in /tmp/main2main/steps/ for detailed changes
4. Optionally push changes to remote branch

• vLLM version: v0.20.2
• vLLM main: vllm-project/vllm@0d4d334